7.3 Extensions to a General Class of Discrete-Time Nonlinear Systems
In this section, we extend the results of the previous sections to a more general class of discrete-time nonlinear systems that may not necessarily be affine in u and w. We consider the general class of systems described by the state-space equations on X⊂ ℜn
∑d : {xk+1 = F(xk,wk,uk); x(k0)=x0zk = Z(xk,wk,uk)yk = Y(xk,wk) |
(7.64) |
where all the variables have their usual meanings, and F : X×W×U → X,F : (0,0,0) = 0,Z : X×W×U → ℜs,Z(0,0,0) = 0, Y: X×W → ℜm
7.3.1 Full-Information H∞
To solve the full-information and state-feedback problems for the system Σd (7.64), we consider the Hamiltonian function:
˜H(x,u,w)=˜V(F(x,w,u))−˜V(x)+12(‖Z(x,u,w)‖2−γ2‖w‖2) |
(7.65) |
for some positive-definite function ˜V : X → ℜ+
∂2˜H∂(u,w)2(0,0,0) = [∂2˜H∂u2∂2˜H∂w∂u∂2˜H∂u∂w ∂2˜H∂w2](0,0,0)
where
huu(0) =[(∂F∂u)T∂2˜V∂λ2(0)(∂F∂u)+(∂Z∂u)T(∂Z∂u)]|x=0,u=0,w=0hww(0) =[(∂F∂w)T∂2˜V∂λ2(0)(∂F∂w)+(∂Z∂w)T(∂Z∂w)−γ2I]|x=0,u=0,w=0huw(0) =[(∂F∂u)T∂2˜V∂λ2(0)(∂F∂w)+(∂Z∂u)T(∂Z∂w)]|x=0,u=0,w=0
Suppose now that the following assumption holds:
(GN1)
huu(0)>0, hww(0)−hwu(0)h−1uu(0)huw(0)<0
Then by the Implicit-function Theorem, the above assumption implies that there exist unique smooth functions ˜u⋆(x),˜w⋆(x)
of x = 0 and satisfying the functional equations
0 = ∂˜H∂u(x,˜u⋆(x),˜w⋆(x)) = (∂˜V∂F∂F∂u+ZT∂Z∂u) |u=˜u⋆(x),w=˜w⋆(x) |
(7.66) |
0 = ∂˜H∂w(x,˜u⋆(x),˜w⋆(x)) =(∂˜V∂F∂F∂w+ZT∂Z∂w) |u=˜u⋆(x),w=˜w⋆(x). |
(7.67) |
Similarly, let
huu(x) ≜ ∂2˜H∂u2(x,˜u⋆(x),˜w⋆(x))hww(x) ≜ ∂2˜H∂w2(x,˜u⋆(x),˜w⋆(x))huw(x) ≜ ∂2˜H∂w∂u(x,˜u⋆(x),˜w⋆(x))=hTwu(x,˜u⋆,˜w⋆)
be associated with the optimal solutions ˜u⋆(x),˜w⋆(x)
(GN2)
˜V(F(x,˜w⋆(x),˜u⋆(x)))−˜V(x)+12(‖Z(x,˜u⋆(x),˜w⋆(x))‖2−γ2‖˜w⋆(x)‖2)=0, ˜V(0)=0. |
(7.68) |
Then we have the following result for the solution of the full-information problem.
Theorem 7.3.1 Consider the discrete-time nonlinear system (7.64), and suppose there exists a C2 positive-definite function ˜V : X1 ⊂ X → ℜ+
(GN3) Any bounded trajectory of the free system
xk+1=F(xk,0,uk),
under the constraint
Z(xk,0,uk)=0
for all k ∈ Z+, is such that, limk→∞ xk = 0.
Then there exists a static full-information feedback control
uk=ˉu⋆(xk)−h−1uu(xk)huw(xk)(wk−ˉw⋆(xk))
which solves the DFIFBNLHICP for the system.
Proof: The proof can be pursued along similar lines as Theorem 7.1.2. □
The above result can also be easily specialized to the state-feedback case as follows.
Theorem 7.3.2 Consider the discrete-time nonlinear system (7.64), and suppose there exists a C2 positive-definite function ˜V : X2 ⊂ X → ℜ+
(GN1s)
huu(0)>0, hww(0)<0
and hypotheses (GN2), (GN3) above. Then, the static state-feedback control
uk=ˉu⋆(xk)
solves the DSFBNLHICP for the system.
Proof: The theorem can be proven along similar lines as Theorem 7.1.3.□
Moreover, the parametrization of all static state-feedback controllers can also be given in the following theorem.
Theorem 7.3.3 Consider the discrete-time nonlinear system (7.64), and suppose the following hypothesis holds
(GN2s) there exists a C2 positive-definite function ˜V : X3 ⊂ X → ℜ+
for some arbitrary smooth function ψ : X3 → ℜp, ψ(0) = 0,
as well as the hypotheses (GN3) and (GN1s) with ˜V
(7.70) |
is a parametrization of all static state-feedback controllers that solves the DSFBNLHICP for the system.
In this subsection, we discuss briefly the output measurement-feedback problem for the general class of nonlinear systems (7.64). Theorem 7.2.1 can easily be generalized to this class of systems. As in the previous case, we can postulate the existence of a dynamic compensator of the form:
∑˜dcdynobs :{ξk+1 = F(ξk,˜w⋆(ξk),˜u⋆(ξk))+˜G(ξk)(yk−Y(ξk,˜w⋆(ξk)) uk = ˜u⋆(ξk) |
(7.71) |
where ˜G
xek+1 = Fe(xek,wk) zk = Ze(xek,wk)
where xe = [xTξT]T,
Fe(xe,w) = [ F(x,w,˜u⋆)F(x,˜w⋆(ξ),˜u⋆(ξ))+˜G(ξ)(y(x,w)−y(ξ,˜w⋆(ξ)))] ,
Ze(xe,w)=Z(x,w,˜u⋆(ξ)).
Then the following result is a direct extension of Theorem 7.2.1.
Theorem 7.3.4 Consider the discrete-time nonlinear system (7.64) and assume the following:
(i) Assumption (GN3) holds and rank κ{∂Z∂u(0,0,0)} = p.
(ii) There exists a C2 positive-definite function ˜V : X → ℜ+locally defined in a neighborhood X of x = 0 satisfying Assumption (GN2).
(iii) There exists an output-injection gain matrix ˜G(.) and a C2 real-valued function W : X4 × X4 locally defined in a neighborhood X4 × X of (x, ξ) = (0, 0), X4 ∩ X4 ≠ ∅, with W (0, 0) = 0, W (x, ξ) > 0 ∀x ≠ ξ and satisfying
(GNM1)
FeT(0,0,0) ∂2W∂xe2(0,0)Fe(0,0,0)+hww(0)<0;
(GNM2)
W(Fe(xe,˜α1(xe)−W(xe)+V(F(x,˜α1(xe)u⋆(ξ)))−V(x)+ 12(‖Z(x,˜α1(xe),˜u⋆(ξ))‖2−γ2‖˜α1(xe)‖2)=0,
where ˜α1(xe) = 0 with ˜α1(0) = 0is a locally unique solution of the equation
∂W∂β|β=Fe(xe,w) ∂Fe∂w(xe,w)+∂V∂λ|λ=F(x,w,˜u⋆(ξ)) ∂F∂w(x,w,˜u⋆(ξ))+ ZT(x,,w,˜u⋆(ξ))∂Z∂w(x,w,˜u⋆(ξ))−γ2wT=0
(GNM3) The discrete-time nonlinear system
xκ+1=F(ξ,˜w⋆(ξ),0)−˜G(ξ)Y(ξ,˜w⋆(ξ))
is locally asymptotically-stable at ξ = 0.
Then, the DMFBNLHICP for the system (7.64) is solvable with the compensator (7.71).
7.4 Approximate Approach to the Discrete-Time Nonlinear H∞-Control Problem
In this section, we discuss alternative approaches to the discrete-time nonlinear H∞-Control problem for affine systems. It should have been observed in Sections 7.1, 7.2, that the control laws that were derived are given implicitly in terms of solutions to certain pairs of algebraic equations. This makes the computational burden in using this design method more intensive. Therefore, in this section, we discuss alternative approaches, although approximate, but which can yield explicit solutions to the problem. We begin with the state-feedback problem.
7.4.1 An Approximate Approach to the Discrete-Time State-Feedback Problem
We consider again the nonlinear system (7.1), and assume the following.
Assumption 7.4.1
rank{k12(x)}=p.
Reconsider now the Hamiltonian function (7.11) associated with the problem:
H2(w,u) = V(f(x)+g1(x)w+g2(x)u)−V(x)+12(‖h1(x)+k11(x)w+k12(x)u‖2− γ2‖w‖2) |
(7.72) |
for some smooth positive-definite function V : X → ℜ+. Suppose there exists a smooth real-valued function ˉu(x)∈ℜp,ˉu(0)=0 such that the HJI-inequality
(7.73) |
is satisfied for all x ∈ X and w ∈ W. Then it is clear from the foregoing that the control law
u=ˉu(x)
solves the DSFBNLHICP for the system Σda globally. The bottleneck however, is in getting an explicit form for the function ū(x). This problem stems from the first term in the HJI-inequality (7.73), i.e.,
V(f(x)+g1(x)w+g2(x)u,
which is a composition of functions and is not necessarily quadratic in u, as in the continuous-time case. Thus, suppose we replace this term by an approximation which is “quadratic” in (w, u), and nonlinear in x, i.e.,
V(f(x)+υ)=V(f(x))+Vx(f(x))υ+12υTVxx(f(x))υ+Rm(x,υ)
for some vector function v ∈ X and where Rm is a remainder term such that
limυ→0 Rm(x,υ)‖υ‖2=0
Then, we can seek a saddle-point for the new Hamiltonian function
ˆH2(w,u) = V(f(x))+Vx(f(x))(g1(x)w+g2(x)u)+ 12(g1(x)w+g2(x)u)TVxx(f(x))(g1(x)+g2(x)u)−V(x)+ 12‖h1(x)+k11(x)w+k12(x)u‖2− 12γ2‖w‖2 |
(7.74) |
by neglecting the higher-order term Rm(x, g1(x)w + g2(x)u). Since ˆH2(u, w) is quadratic in (w, u), it can be represented as
ˆH2(w,u) = V(f(x))−V(x)+ 12hT1(x)h1(x)+ˆS(x)[wu]+12[wu]ˆR(x)[wu],
where
ˆS(x)=hT1(x)[k11(x) k12(x)]+Vx(f(x))[g1(x) g2(x)]
and
ˆR(x)=(kT11(x) k11(x)−γ2I kT11(x) k12(x) kT12(x) k11(x) kT12(x) k12(x) )+(gT1(x)gT2(x))Vxx(f(x))(g1(x) g2(x))
From this, it is easy to determine conditions for the existence of a unique saddle-point and explicit formulas for the coordinates of this point. It can immediately be determined that, if ˆR(x) is nonsingular, then
ˆH2(w,u)=ˆH2(w⋆(x),u⋆(x))+12[w−ˆw⋆(x) u−ˆu⋆(x)]TˆR(x)[w−ˆw⋆(x) u−ˆu⋆(x)] |
(7.75) |
where
(7.76) |
One condition that gurantees that ˆR(x) is nonsingular (by Assumption 7.4.1) is that the submatrix
ˆR11(x):=k11T(x)k11(x)−γ2I+1/2g1T(x)Vxx(f(x))g1(x)<0 ∀x.
If the above condition is satisfied for some γ > 0, then ˆH2(x,w,u) has a saddle-point at (û, ŵ ), and
ˆH2(ˆw⋆(x),ˆu⋆(x))=V(f(x))−V(x)−12ˆS(x)ˆR−1(x)ˆST(x)+12hT1(x)h1(x). |
(7.77) |
The above development can now be summarized in the following lemma.
Lemma 7.4.1 Consider the discrete-time nonlinear system (7.1), and suppose there exists a smooth positive-definite function V : X0 ⊂ X → +, V (0) = 0 and a positive number δ > 0 such that
(i)
(7.78) |
(ii)
ˆH2(ˆw⋆(x),ˆu⋆(x))<−12δ(||ˆw⋆(x)||2+||ˆu⋆(x)||2) ∀x∈X0,
(iii)
ˆR11(x)<0 ∀x∈X0.
Then, there exists a neighborhood X × W of (w, x) = (0, 0) in X×W such that V satisfies the HJI-inequality (7.73) with ū = û⋆ (x), û⋆ (0) = 0.
Proof: By construction, H2(x, w, u) satisfies
ˆH2(x,w,u)=ˆH2(ˆw⋆(x),ˆu⋆(x))+12(w−ˆw⋆)TˆR11(x)(w−ˆw⋆(x)).
Since R11(0) is negative-definite by hypothesis (iii), there exists a neighborhood X1 of x = 0 and a positive number c > 0 such that
(w−ˆw⋆)TˆR11(x)(w−ˆw⋆(x))≤c‖(w−ˆw⋆(x)‖2 ∀x∈X1, ∀w.
Now let μ = min{δ, c}
(w−ˆw⋆)TˆR11(x)(w−ˆw⋆(x))≤μ‖(w−ˆw⋆(x)‖2 ∀x∈X1.
Moreover, by hypothesis (ii)
ˆH2(ˆw⋆(x),ˆu⋆(x))≤−μ2(‖ˆw⋆(x)‖2+‖ˆu⋆(x)‖2) ∀x∈X0.
Thus, by the triangle inequality,
ˆH2(w,ˆu⋆(x)) ≤ −μ2(‖ˆw⋆(x)‖2+12‖ˆu⋆(x)‖2+12‖(w−ˆw⋆(x)‖2) ≤ −μ2(‖w‖2+‖ˆu*(x)‖2) |
(7.79) |
for all x ∈ X2, where X2 = X0 ∩ X1. Notice however that the Hamiltonians H2(x, w, u,) and Ĥ2(x, w, u) defined by (7.72) and (7.74) respectively, are related by
(7.80) |
By the result in Section 8.14.3 of reference [94], for all κ > 0, there exist neighborhoods X3 of x = 0, W1 of w = 0 and U1 of u = 0 such that
(7.81) |
Finally, combining (7.79), (7.80) and (7.81), one obtains an estimate for H2(w, u (x)), i.e.,
H2(w,u⋆(x))≤−μ2(1−κμ)(‖w‖2+‖ˆu⋆(x)‖2)
Choosing κ < μ and X ⊆ X3, W1 ⊆ W, the result follows. □
From the above lemma, one can conclude the following.
Theorem 7.4.1 Consider the discrete-time nonlinear system (7.1), and assume all the hypotheses (i), (ii), (iii) of Lemma 7.4.1 hold. Then the closed-loop system Σda :
∑da : {xk+1 = f(xk)+g1(xk)wk+g2(xk)ˉu(xk); x0=0zk = h1(xk)+k11(xk)wk+k12(xk)ˉu(xk) |
(7.82) |
with ū(x) = û (x) has a locally asymptotically-stable equilibrium-point at x = 0, and for every K ∈ Z+, there exists a number > 0 such that the response of the system from the initial state x0 = 0 satisfies
K∑k=0‖zk‖2≤γ2K∑k=0‖wk‖2
for every sequence w = (w0,…, wK) such that wk < ε.
Proof: Since f(.) and û (.) are smooth and vanish at x = 0, it is easily seen that for every K > 0, there exists a number > 0 such that the response of the closed-loop system to any input sequence w = (w0,…, wK) from the initial state x0 = 0 is such that xk ∈ X for all k ≤ K + 1 as long as wk < for all k ≤ K. Without any loss of generality, we may assume that is such that wk ∈ W. In this case, using Lemma 7.4.1, we can deduce that the dissipation-inequality
V(xk+1)−V(xk)+12(zTkzk−γ2|wTkwk)≤0 ∀k≤K
holds. The result now follows from Chapter 3 and a Lyapunov argument. □
Remark 7.4.1 Again, in the case of the linear system Σdl (7.17), the result of Theorem 7.4.1 reduces to wellknown necessary and sufficient conditions for the existence of a solution to the linear DSFBNLHICP [89]. Indeed, setting B := [B1 B2] and D := [D11 D12], a quadratic function V (x) = 12 xT P x with P = P T > 0 satisfies the hypotheses of the theorem if and only if
ATPA−P+CT1C1−FTp(R+BTPB)Fp < 0 DT11D11−γ2I+BT1PB1 < 0
where Fp=−(R+BTPB)−1(BTPA+DTC1).. In this case,
[ˆw*ˆu*]=Fpx.
In the next subsection, we consider an approximate approach to the measurement-feedback problem.
7.4.2 An Approximate Approach to the Discrete-Time Output Measurement-Feedback Problem
In this section, we discuss an alternative approximate approach to the discrete-time measurement-feedback problem for affine systems. In this regard, assume similarly a dynamic observer-based controller of the form
∑ˉdacdynobs : {θk+1 = f(θk)+g1(θk) ˆw⋆(θk)+g2(θk) ˆu⋆(θk)+ˉG(θk)[yk−h2(θk)− k21(θk)ˆw⋆(θk) uk = ˆu⋆(θk) |
(7.83) |
where θ ∈ X is the controller state vector, while ŵ (.), û (.) are the optimal state-feedback control and worst-case disturbance given by (7.76) respectively, and ˉG is the output-injection gain matrix which is to be determined. Accordingly, the corresponding closed-loop system (7.1), (7.83) can be represented by
(7.84) |
where x# = [xT θT ]T ,
f#(x#) = [f(x)+g2(x)ˆu⋆(θ)(f(θ)+g1(θ)ˆw⋆(θ)+g2(θ)ˆu⋆(θ)+ˉG(θ)(h2(x)−h2(θ)−k21(θ)ˆw⋆(θ))],
g#(x#) = [g1(x)ˉG(θ)k21(x)],
and
h#(x#) = h1(x)+k12(x)ˆu⋆(θ), k#(x#)=k11(x).
The objective is to find sufficient conditions under which the above closed-loop system (7.84) is locally (globally) asymptotically-stable and the estimate θ → x as t → ∞. This can be achieved by first rendering the closed-loop system dissipative, and then using some suitable additional conditions to conclude asymptotic-stability.
Thus, we look for a suitable positive-definite function Ψ1 : X × X → +, such that the dissipation-inequality
(7.85) |
is satisfied along the trajectories of the system (7.84). To achieve this, we proceed as in the continuous-time case, Chapter 5, and assume the existence of a smooth C2 function W : X × X → such that W (x ) ≥ 0 for all x = 0 and W (x ) > 0 for all x = θ. Further, set
H#2(w)=W(f#(x#)+g#(x#)w)−W(x#)+H2(w,ˆu⋆(θ))−ˆH2(ˆw⋆(x)ˆu⋆(x)), |
(7.86) |
and recall that
H2(w,u)=ˆH2(w,u)+Rm(x,g1(x)w+g2(x)u);
so that
H2(w,ˆu⋆(θ))=ˆH2(w,ˆu⋆(θ))+Rm(x,g1(x)w+g2(x)ˆu⋆(x)).
Moreover, by definition
H2(w,ˆu⋆(θ)) = V(f(x)+g1(x)w+g2(x)ˆu⋆(θ)−V(x)+12||h2(x)+ k11(x)w+ +k12(x)ˆu⋆(θ)||2−12γ2||w||2. |
(7.87) |
Therefore, subsituting (7.87) in (7.86) and rearranging, we get
H#2(w)+ˆH2(ˆu⋆(x),ˆw⋆(x)) = W(f#(x#)+g#(x#)w)−W(x#)+V(f(x)+g1(x)w+g2(x)ˆu⋆(θ))−V(x)+12||h#(x#)+k#(x#)w||2−12γ2||w||2 |
(7.88) |
From the above identity (7.88), it is clear that, if the right-hand-side is nonpositive, then the positive-definite function
Ψ1(x#)=W(x#)+V(x)
will indeed have satisfied the dissipation-inequality (7.85) along the trajectories of the closed-loop system (7.84). Moreover, if we assume the hypotheses of Theorem 7.4.1 (respectively Lemma 7.4.1) hold, then the term 5H2(û (x), ŵ (x)) is nonpositive. Therefore, it remains to impose on H#2(w) to be also nonpositive for all w.
One way to achieve the above objective, is to impose the condition that
maxwH2#(w)<0.
However, finding a closed-form expression for w⋆⋆ = arg max{H2(w)} is in general not possible as observed in the previous section. Thus, we again resort to an approximate but practical approach. Accordingly, we can replace the term W (f (x ) + v ) in H2(.) by its second-order Taylor approximation as:
W(f#(x#)+υ#)=W(f#(x#))+Wx#(f#(x#))υ#+12υ#TWx#x#(f#(x#))υ#+R#m(x#,υ#)
for any v ∈ X × X, and where Rm(x , v ) is the remainder term. While the last term (recalling from equation (7.75)) can be represented as
H2(w,ˆu⋆(θ)) −ˆH2(ˆw⋆(x),ˆu⋆(x)) = 12[ w−ˆw⋆(x)ˆu⋆(θ)−ˆu⋆(x)]TˆR(x)[ w−ˆw⋆(x)ˆu⋆(θ)−ˆu⋆(x)] +Rm(x,g1(x)w+g2(x)ˆu⋆(θ)).
Similarly, observe that Rm(x, g1(x)w + g2(x)û⋆ (θ)) can be expanded as a function of x# with respect to w as
Rm(x,g1(x)w+g2(x)ˆu⋆(θ))=Rm0(x#)+Rm1(x#)w+wTRm2(x#)w+Rm3(x#,w).
Thus, we can now approximate H2(w) with the function
H#2(w) = W(f#(x#))−W(x#)+ Rm0(x#)+Rm1(x#)+Wx#(f#(x#))g#(x#)w+ wT(Rm2(x#)+12g#T(x#)Wx#x#(f#(x#))g#(x#))w+ 12[ w−ˆw⋆(x)ˆu⋆(θ)−ˆu⋆(x)]TˆR(x)[ w−ˆw⋆(x)ˆu⋆(θ)−ˆu⋆(x)] . |
(7.89) |
Moreover, we can now determine an estimate ŵ of w from the above expression (7.89) for H2(w) by taking derivatives with respect to w and solving the linear equation
∂ˉH#2∂w(ˆw⋆⋆) =0.
It can be shown that, if the matrix
ˉR(x#)=12g#(x#)TWx#x#(f#(x#))g#(x#)+ˆR11(x)+Rm2(x#)
is nonsingular and negative-definite, then ŵ is unique, and is a maximum for H2(w). The design procedure outlined above can now be summarized in the following theorem.
Theorem 7.4.2 Consider the nonlinear discrete-time system (7.1) and suppose the following hold:
(i) there exists a smooth positive-definite function V defined on a neighborhood X0 ⊂ X of x = 0, satisfying the hypotheses of Lemma 7.4.1.
(ii) there exists a smooth positive-semidefinite function W (x ), defined on a neighborhood Ξ of x = 0 in X×X, such that W (x ) > 0 for all x, x = θ, and satisfying
ˉH#2(ˆw⋆⋆(x#))<0
for all 0 = x ∈ Ξ. Moreover, there exists a number δ > 0 such that
ˉH#2(ˆw⋆⋆(x#))<−δ||ˆw⋆⋆(x#)||2 ˉR(x#) <0
for all x ∈ Ξ.
Then, the controllerˉΣdacdynobsdynobs given by (7.83) locally asymptotically stabilizes the closed-loop system (7.84), and for every K ∈ Z+, there is a number ε > 0 such that the response from the initial state (x0, θ0) = (0, 0) satisfies
K∑k=0zTkzk ≤ γ2K∑k=0wTkwk
for every sequence w = (w0,…, wK) such that wk2 <ε.
This chapter is entirely based on the papers by Lin and Byrnes [182]-[185]. In particular, the discussion on controller parameterization is from [184]. The results for stable plants can also be found in [51]. Similarly, the results for sampled-data systems have not been discussed here, but can be found in [124, 213, 255].
The alternative and approximate approach for solving the discrete-time problems is mainly from Reference [126], and approximate approaches for solving the DHJIE can also be found in [125]. An information approach to the discrete-time problem can be found in [150], and connections to risk-sensitive control in [151, 150].